Tensorflow.js tf.train.adam() Function

Last Updated : 21 Nov, 2022

Tensorflow.js is a javascript library developed by Google to run and train machine learning model in the browser or in Node.js.

Adam optimizer (or Adaptive Moment Estimation) is a stochastic gradient descent method that is based on adaptive estimation of first-order and second-order moments. The optimization technique is highly efficient in when working with a large sets of data and parameters. For more details refer to this article.

In Tensorflow.js tf.train.adam() function is used which creates tf.AdamOptimizer that uses the adam algorithm.

Syntax:

tf.train.adam (learningRate?, beta1?, beta2?, epsilon?)

Parameters:

learningRate: The learning rate to use for the Adam gradient descent algorithm. It is optional.
beta1: The exponential decay rate for the 1st moment estimates. It is optional.
beta2: The exponential decay rate for the 2nd moment estimates. It is optional.
epsilon: A small constant for numerical stability. It is optional.

Return Value: AdamOptimizer.

Example 1: A quadratic function is defined with taking x, y input tensors and a, b, c as random coefficients. Then we calculate the mean squared loss of the prediction and pass it to adam optimizer to minimize the loss with and change the coefficient ideally.

Javascript

// A cubic function with its coefficient l,m,n.
const x = tf.tensor1d([0, 1, 2, 3]);
const y = tf.tensor1d([1., 2., 5., 11.]);
 
const l = tf.scalar(Math.random()).variable();
const m = tf.scalar(Math.random()).variable();
const n = tf.scalar(Math.random()).variable();
 
// y = l * x^3 - m * x + n.
const f = x => l.mul(x.pow(3)).sub(m.mul(x)).add(n);
const loss = (pred, label) => pred.sub(label).square().mean();
 
const learningRate = 0.01;
const optimizer = tf.train.adam(learningRate);
 
// Training the model and printing the coefficients.
for (let i = 0; i < 10; i++) {
   optimizer.minimize(() => loss(f(x), y));
  console.log(
     `l: ${l.dataSync()}, m: ${m.dataSync()}, n: ${n.dataSync()}`);
}
 
// Predictions are made.
 
const preds = f(x).dataSync();
preds.forEach((pred, i) => {
   console.log(`x: ${i}, pred: ${pred}`);
});

Output:

l: 0.5212615132331848, m: 0.4882013201713562, n: 0.9879841804504395
l: 0.5113212466239929, m: 0.49809587001800537, n: 0.9783468246459961
l: 0.5014950633049011, m: 0.5077731013298035, n: 0.969675600528717
l: 0.49185076355934143, m: 0.5170749425888062, n: 0.9630305171012878
l: 0.48247095942497253, m: 0.5257879495620728, n: 0.9595866799354553
l: 0.47345229983329773, m: 0.5336435437202454, n: 0.9596782922744751
l: 0.4649032950401306, m: 0.5403363704681396, n: 0.9626657962799072
l: 0.4569399356842041, m: 0.5455683469772339, n: 0.9677067995071411
l: 0.4496782124042511, m: 0.5491118431091309, n: 0.9741682410240173
l: 0.44322386384010315, m: 0.5508641004562378, n: 0.9816395044326782
x: 0, pred: 0.9816395044326782
x: 1, pred: 0.8739992380142212
x: 2, pred: 3.4257020950317383
x: 3, pred: 11.29609203338623

Example 2: Below is the code where we designed a simple model and we define an optimizer by tf.train.adam with the learning rate parameter of 0.01 and pass it to model compilation.

Javascript

// Importing the tensorflow.js library
import * as tf from "@tensorflow/tfjs"
 
// defining the model
const model = tf.sequential({
    layers: [tf.layers.dense({ units: 1, inputShape: [12] })],
});
   
 
// in the compilation we use tf.train.adam optimizer 
  const optimizer = tf.train.adam(0.001);
model.compile({ optimizer: optimizer, loss: "meanSquaredError" },
              (metrics = ["accuracy"]));
   
// evaluate the model which was compiled above
const result = model.evaluate(tf.ones([10, 12]), tf.ones([10, 1]), {
    batchSize: 4,
});
   
// print the result
result.print();